Medical and Biomedical Imaging

The Image Team has an expertise in Medical and Biomedica Imaging. Since 2017, the team worked on various topics such as lesion detection in brain or cardiac MRIs, brain structure segmentation (neonate, isointense phase or adult), survival prediction…

Our methods uses not only Deep Learning, but also traditionnal image processing techniques, especially mathematical morphology.

Main contact: Elodie Puybareau

Projects and collaborations

Current

• Elodie Puybareau and Thierry Géraud co-supervise a thesis with GE Healthcare on the detection of anomalies in pulmonary vessels using few shot learning. More information here.

• We collaborate with the IGR (Institut Gustave Roussy, a French reference center for cancer) to work on radiotherapies. The generation of radiotherapies for the treatment of human cancer is a particularly expensive and time-consuming task in the medical community. We propose to use an unsupervised approach to generate schedules for treatments based on different modalities (IMRT, VMAT, 3D, …).

Past

• Collaborations:

  • Development of a software dedicated to the segmentation of brain structures and DEHSI in neonates MRI, in collaboration with LTCI (Télécom ParisTech), Kremlin-Bicetre Hospital, Tours Hospital. More information here.

• Former PhD Student:

  • Zhao Zhou, under the supervision of Thierry Géraud and Elodie Puybareau, “Heart segmentation and evaluation of fibrosis”, defended 10th of January 2023.

Main results

Light and Fast CNN

One of our main contribution is to provide fast and robust pipelines. We hence considered to use VGG16 to segment 3D medical images, using 3 slices concatened in 2D-color input images. This techniques can be use with consecutive slices to take advantage of the 3D information, or slices coming from different modalities, or a mix between slices and modalities as shown in the figure above, on BraTS dataset.

Challenges

The Image Team also participate in challenges and obtained interesting results since 2017 with our first participation to iSeg and WHM where we obtained good results (in the top 3 methods for some metrics). For the “officials” rankings, we can highlight the following challenges:

Topics of interest

• Segmentation
• Lesion detection
• Prediction
• Fast and light procedures